Domain Adaptation In Reinforcement Learning Via Latent Unified State Representation